Single-Loop Byzantine-Resilient Federated Bilevel Optimization

摘要

Federated bilevel optimization plays a crucial role in solving complex problems with nested optimization structures. However, its distributed nature makes it highly susceptible to faulty or Byzantine behaviors. Existing Byzantine-resilient approaches are either restricted to simple single-level optimization problems or rely on sub-loop updates that introduce significant computational and communication overhead. To address these limitations, we propose a family of Byzantine-resilient federated bilevel algorithms, which (i) operate within a single-loop structure, (ii) achieve optimal Byzantine resilience, and (iii) ensure computational and communication efficiency. The core of the proposed method, BR-FedBi, leverages an auxiliary variable that facilitates efficient hypergradient estimation while simultaneously solving the lower- and upper-level problems. Building on BR-FedBi, we further integrate the algorithm with Polyak’s momentum and the probabilistic gradient estimator (PAGE) (Li et al., 2021), resulting in provable optimal Byzantine resilience and optimal sample complexity. Both theoretical analysis and empirical results demonstrate the superior performance of the proposed algorithms.